base.py 28.1 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import asyncio
4
import copy
5
import time
6
from abc import ABC, abstractmethod
7
from collections.abc import Mapping, Sequence
8
9
10
11
from functools import cached_property
from typing import TYPE_CHECKING, Any, Generic, overload

from typing_extensions import TypeVar
12

13
14
15
16
17
18
19
20
21
22
23
from vllm.inputs import (
    EmbedsInputs,
    EmbedsPrompt,
    EncoderDecoderInputs,
    ProcessorInputs,
    SingletonInputs,
    TextPrompt,
    TokenInputs,
    TokensPrompt,
)
from vllm.inputs.data import build_enc_dec_inputs, embeds_inputs, token_inputs
24
from vllm.logger import init_logger
25
from vllm.tokenizers import TokenizerLike
26
from vllm.utils.async_utils import AsyncMicrobatchTokenizer
27
from vllm.utils.counter import AtomicCounter
28
29
from vllm.utils.torch_utils import set_default_torch_num_threads
from vllm.v1.metrics.stats import MultiModalCacheStats
30
31

from .embed_utils import safe_load_prompt_embeds
32
33
34
35
from .inputs import (
    DictPrompt,
    EncoderDecoderDictPrompt,
    EncoderDecoderTokPrompt,
36
37
    SingletonDictPrompt,
    SingletonTokPrompt,
38
39
    TokPrompt,
)
40
from .inputs.preprocess import extract_target_prompt
41
from .params import ChatParams, TokenizeParams
42
43

if TYPE_CHECKING:
44
    from vllm.config import VllmConfig
45
46
47
48
    from vllm.entrypoints.chat_utils import (
        ChatCompletionMessageParam,
        ConversationMessage,
    )
49
    from vllm.multimodal.cache import BaseMultiModalProcessorCache
50
51
52
53
54
    from vllm.multimodal.inputs import (
        MultiModalDataDict,
        MultiModalInputs,
        MultiModalUUIDDict,
    )
55
    from vllm.multimodal.parse import MultiModalDataItems, MultiModalUUIDItems
56
    from vllm.multimodal.processing import BaseMultiModalProcessor
57

58
59
logger = init_logger(__name__)

60

61
62
63
64
_T = TypeVar("_T", bound=TokenizerLike, default=TokenizerLike)


class BaseRenderer(ABC, Generic[_T]):
65
    @classmethod
66
    @abstractmethod
67
68
    def from_config(
        cls,
69
        config: "VllmConfig",
70
        tokenizer_kwargs: dict[str, Any],
71
    ) -> "BaseRenderer":
72
73
        raise NotImplementedError

74
    def __init__(self, config: "VllmConfig", tokenizer: _T | None) -> None:
75
76
        super().__init__()

77
        self.config = config
78
        self.model_config = config.model_config
79
        self.api_process_rank = config.parallel_config._api_process_rank
80

81
82
        self.tokenizer = tokenizer

83
84
85
        # Lazy initialization since offline LLM doesn't use async
        self._async_tokenizer: AsyncMicrobatchTokenizer | None = None

86
87
88
89
        self.mm_processor: BaseMultiModalProcessor | None = None
        self._mm_cache_stats: MultiModalCacheStats | None = None
        if config.model_config.is_multimodal_model:
            from vllm.multimodal import MULTIMODAL_REGISTRY as mm_registry
90
            from vllm.multimodal.registry import MultiModalTimingRegistry
91
92
93

            mm_processor_cache = mm_registry.processor_cache_from_config(config)

94
95
96
97
98
99
100
            # Deep-copy the tokenizer so the multimodal processor gets its
            # own Rust tokenizer backend.  Without this, concurrent access
            # from AsyncMicrobatchTokenizer and call_hf_processor causes
            # "RuntimeError: Already borrowed" from the Rust RefCell.
            # See: https://github.com/huggingface/tokenizers/issues/537
            mm_tokenizer = copy.deepcopy(tokenizer)

101
102
103
            with set_default_torch_num_threads():
                self.mm_processor = mm_registry.create_processor(
                    config.model_config,
104
                    tokenizer=mm_tokenizer,
105
106
107
108
109
                    cache=mm_processor_cache,
                )

            if mm_processor_cache:
                self._mm_cache_stats = MultiModalCacheStats()
110

111
112
113
            # This is used to generate internal request ID for MM processing
            # It has no relation to the request ID for engine core
            self._mm_req_counter = AtomicCounter()
114
115
116
            self._mm_timing_registry = MultiModalTimingRegistry(
                config.observability_config
            )
117

118
    def get_tokenizer(self) -> _T:
119
120
121
122
123
124
        tokenizer = self.tokenizer
        if tokenizer is None:
            raise ValueError("Tokenizer not available when `skip_tokenizer_init=True`")

        return tokenizer

125
    def get_async_tokenizer(self) -> AsyncMicrobatchTokenizer:
126
        if self._async_tokenizer is None:
127
128
129
130
            self._async_tokenizer = AsyncMicrobatchTokenizer(self.get_tokenizer())

        return self._async_tokenizer

131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
    def get_mm_processor(self) -> "BaseMultiModalProcessor":
        if self.mm_processor is None:
            raise ValueError("Multi-modal processor not available for text-only models")

        return self.mm_processor

    @property
    def mm_processor_cache(self) -> "BaseMultiModalProcessorCache | None":
        if self.mm_processor is None:
            return None

        return self.mm_processor.cache

    def stat_mm_cache(self) -> MultiModalCacheStats | None:
        mm_cache_stats = self._mm_cache_stats
        if mm_cache_stats is None:
            return None

        self._mm_cache_stats = MultiModalCacheStats()

        return mm_cache_stats

    def update_mm_cache_stats(self) -> None:
        mm_processor_cache = self.mm_processor_cache
        mm_cache_stats = self._mm_cache_stats

        if mm_processor_cache and mm_cache_stats:
            delta = mm_processor_cache.make_stats(delta=True)
            mm_cache_stats.record(delta.total, delta.hits)

    def clear_mm_cache(self) -> None:
        mm_processor_cache = self.mm_processor_cache
        if mm_processor_cache is not None:
            mm_processor_cache.clear_cache()

        if self._mm_cache_stats is not None:
            self._mm_cache_stats.reset = True

169
170
171
172
173
174
175
176
177
178
    def warmup(self, chat_params: ChatParams) -> None:
        """
        Warm up this renderer to avoid first-request latency.

        For chat requests:
        - Jinja2 template compilation

        For multi-modal requests:
        - Importing libraries such as librosa triggers JIT compilation.
        """
179
180
        from vllm.entrypoints.chat_utils import ChatTemplateResolutionError

181
        try:
182
            logger.debug("Warming up chat template processing...")
183
184
185
186
187
            start_time = time.perf_counter()

            self.render_chat([[{"role": "user", "content": "warmup"}]], chat_params)

            elapsed = time.perf_counter() - start_time
188
            logger.debug("Chat template warmup completed in %.3fs", elapsed)
189
        except ChatTemplateResolutionError:
190
            logger.debug("This model does not support chat template.")
191
        except Exception:
192
            logger.warning("Chat template warmup failed", exc_info=True)
193
194
195
196
197
198
199
200
201
202

        if self.mm_processor:
            from vllm.multimodal.processing import TimingContext

            model_config = self.model_config
            mm_config = model_config.get_multimodal_config()
            processor = self.mm_processor
            mm_limits = processor.info.allowed_mm_limits

            try:
203
                logger.debug("Warming up multi-modal processing...")
204
205
206
207
208
209
210
211
                start_time = time.perf_counter()

                processor_inputs = processor.dummy_inputs.get_dummy_processor_inputs(
                    seq_len=model_config.max_model_len,
                    mm_counts=dict.fromkeys(mm_limits, 1),
                    mm_options=mm_config.limit_per_prompt,
                )
                _ = processor.apply(
212
                    processor_inputs, timing_ctx=TimingContext(enabled=False)
213
214
215
216
217
                )

                elapsed = time.perf_counter() - start_time
                logger.info("Multi-modal warmup completed in %.3fs", elapsed)
            except Exception:
218
                logger.warning("Multi-modal warmup failed")
219
220
221
            finally:
                self.clear_mm_cache()

222
223
224
225
226
    def shutdown(self) -> None:
        mm_processor_cache = self.mm_processor_cache
        if mm_processor_cache is not None:
            mm_processor_cache.close()

227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
    def get_bos_token_id(self) -> int | None:
        if self.tokenizer is None:
            logger.warning_once(
                "Using None for BOS token id because tokenizer is not initialized"
            )
            return None

        return self.tokenizer.bos_token_id

    def get_eos_token_id(self) -> int | None:
        if self.tokenizer is None:
            logger.warning_once(
                "Using None for EOS token id because tokenizer is not initialized"
            )
            return None

        return self.tokenizer.eos_token_id

245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
    def get_dec_start_token_id(self) -> int:
        """
        Obtain the decoder start token id employed by an encoder/decoder model,
        raising an error if it is not available.
        """
        dec_start_token_id = getattr(
            self.model_config.hf_config, "decoder_start_token_id", None
        )

        if dec_start_token_id is None:
            logger.warning_once(
                "Falling back on <BOS> for decoder start token id "
                "because decoder start token id is not available."
            )
            dec_start_token_id = self.get_bos_token_id()

        if dec_start_token_id is None:
            raise RuntimeError("Cannot find decoder start token id or <BOS>")

        return dec_start_token_id

266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
    @cached_property
    def default_cmpl_tok_params(self) -> TokenizeParams:
        mm_processor = self.mm_processor
        if mm_processor is not None:
            return mm_processor.info.default_tok_params

        model_config = self.model_config
        encoder_config = model_config.encoder_config or {}

        return TokenizeParams(
            max_total_tokens=model_config.max_model_len,
            do_lower_case=encoder_config.get("do_lower_case", False),
            add_special_tokens=True,
        )

    @cached_property
    def default_chat_tok_params(self) -> TokenizeParams:
        mm_processor = self.mm_processor
        if mm_processor is not None:
            return mm_processor.info.default_tok_params

        model_config = self.model_config
        encoder_config = model_config.encoder_config or {}

        return TokenizeParams(
            max_total_tokens=model_config.max_model_len,
            do_lower_case=encoder_config.get("do_lower_case", False),
            add_special_tokens=False,
        )

296
    # Step 1: Convert raw inputs to prompts
297
    def render_prompt(
298
        self,
299
300
301
        prompt: DictPrompt | bytes,
    ) -> DictPrompt:
        if isinstance(prompt, bytes):
302
            embeds = safe_load_prompt_embeds(self.model_config, prompt)
303
            prompt = EmbedsPrompt(prompt_embeds=embeds)
304

305
        return prompt
306

307
    def render_prompts(
308
        self,
309
310
311
        prompts: Sequence[DictPrompt | bytes],
    ) -> list[DictPrompt]:
        if len(prompts) == 0:
312
313
            raise ValueError("You must pass at least one prompt")

314
        return [self.render_prompt(prompt) for prompt in prompts]
315

316
    async def render_prompts_async(
317
        self,
318
319
320
        prompts: Sequence[DictPrompt | bytes],
    ) -> list[DictPrompt]:
        return self.render_prompts(prompts)
321

322
    @abstractmethod
323
324
325
    def render_messages(
        self,
        messages: list["ChatCompletionMessageParam"],
326
        params: ChatParams,
327
    ) -> tuple[list["ConversationMessage"], DictPrompt]:
328
329
330
331
332
        raise NotImplementedError

    async def render_messages_async(
        self,
        messages: list["ChatCompletionMessageParam"],
333
        params: ChatParams,
334
    ) -> tuple[list["ConversationMessage"], DictPrompt]:
335
336
337
        return self.render_messages(messages, params)

    # Step 2: Tokenize prompts if necessary
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
    def _tokenize_prompt(
        self,
        prompt: TextPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt:
        tokenizer = self.get_tokenizer()
        prompt_token_ids = tokenizer.encode(
            prompt["prompt"],
            **params.get_encode_kwargs(),
        )

        return TokensPrompt(prompt_token_ids=prompt_token_ids, **prompt)

    async def _tokenize_prompt_async(
        self,
        prompt: TextPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt:
        tokenizer = self.get_async_tokenizer()
        prompt_token_ids = await tokenizer.encode(
            prompt["prompt"],
            **params.get_encode_kwargs(),
        )

        return TokensPrompt(prompt_token_ids=prompt_token_ids, **prompt)

    def _detokenize_prompt(self, prompt: TokensPrompt) -> TokensPrompt:
        tokenizer = self.get_tokenizer()
        prompt["prompt"] = tokenizer.decode(prompt["prompt_token_ids"])

        return prompt

    async def _detokenize_prompt_async(self, prompt: TokensPrompt) -> TokensPrompt:
        tokenizer = self.get_async_tokenizer()
        prompt["prompt"] = await tokenizer.decode(prompt["prompt_token_ids"])

        return prompt

376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
    @overload
    def _tokenize_singleton_prompt(
        self,
        prompt: TextPrompt | TokensPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt: ...

    @overload
    def _tokenize_singleton_prompt(  # type: ignore[misc]
        self,
        prompt: EmbedsPrompt,
        params: TokenizeParams,
    ) -> EmbedsPrompt: ...

    def _tokenize_singleton_prompt(
        self,
        prompt: SingletonDictPrompt,
        params: TokenizeParams,
    ) -> SingletonTokPrompt:
        if "prompt_token_ids" not in prompt and "prompt_embeds" not in prompt:
            prompt = params.apply_pre_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
            prompt = self._tokenize_prompt(prompt, params)

        if params.needs_detokenization and "prompt" not in prompt:
            if "prompt_token_ids" not in prompt:
                raise RuntimeError("Cannot run detokenization on embeddings")

            prompt = self._detokenize_prompt(prompt)  # type: ignore[arg-type]

        return params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]

    @overload
    async def _tokenize_singleton_prompt_async(
        self,
        prompt: TextPrompt | TokensPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt: ...

    @overload
    async def _tokenize_singleton_prompt_async(  # type: ignore[misc]
        self,
        prompt: EmbedsPrompt,
        params: TokenizeParams,
    ) -> EmbedsPrompt: ...

    async def _tokenize_singleton_prompt_async(
        self,
        prompt: SingletonDictPrompt,
        params: TokenizeParams,
    ) -> SingletonTokPrompt:
        if "prompt_token_ids" not in prompt and "prompt_embeds" not in prompt:
            prompt = params.apply_pre_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
            prompt = await self._tokenize_prompt_async(prompt, params)

        if params.needs_detokenization and "prompt" not in prompt:
            if "prompt_token_ids" not in prompt:
                raise RuntimeError("Cannot run detokenization on embeddings")

            prompt = await self._detokenize_prompt_async(prompt)  # type: ignore[arg-type]

        return params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]

438
439
440
441
442
443
    def _tokenize_enc_dec_prompt(
        self,
        prompt: EncoderDecoderDictPrompt,
        params: TokenizeParams,
    ) -> EncoderDecoderTokPrompt:
        enc_prompt, dec_prompt = (
444
            self._tokenize_singleton_prompt(prompt["encoder_prompt"], params),
445
446
447
            (
                None
                if prompt["decoder_prompt"] is None
448
                else self._tokenize_singleton_prompt(prompt["decoder_prompt"], params)
449
450
451
452
453
454
455
456
457
458
459
460
461
462
            ),
        )

        return EncoderDecoderTokPrompt(
            encoder_prompt=enc_prompt,
            decoder_prompt=dec_prompt,
        )

    async def _tokenize_enc_dec_prompt_async(
        self,
        prompt: EncoderDecoderDictPrompt,
        params: TokenizeParams,
    ) -> EncoderDecoderTokPrompt:
        enc_prompt, dec_prompt = await asyncio.gather(
463
            self._tokenize_singleton_prompt_async(prompt["encoder_prompt"], params),
464
465
466
            (
                asyncio.sleep(0)
                if prompt["decoder_prompt"] is None
467
468
469
                else self._tokenize_singleton_prompt_async(
                    prompt["decoder_prompt"], params
                )
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
            ),
        )

        return EncoderDecoderTokPrompt(
            encoder_prompt=enc_prompt,
            decoder_prompt=dec_prompt,
        )

    def tokenize_prompt(
        self,
        prompt: DictPrompt,
        params: TokenizeParams,
    ) -> TokPrompt:
        if "encoder_prompt" in prompt:
            return self._tokenize_enc_dec_prompt(prompt, params)  # type: ignore[arg-type]

486
        return self._tokenize_singleton_prompt(prompt, params)
487
488
489

    def tokenize_prompts(
        self,
490
        prompts: Sequence[DictPrompt],
491
        params: TokenizeParams,
492
    ) -> list[TokPrompt]:
493
494
        return [self.tokenize_prompt(prompt, params) for prompt in prompts]

495
496
497
498
499
500
501
    async def tokenize_prompt_async(
        self,
        prompt: DictPrompt,
        params: TokenizeParams,
    ) -> TokPrompt:
        if "encoder_prompt" in prompt:
            return await self._tokenize_enc_dec_prompt_async(prompt, params)  # type: ignore[arg-type]
502

503
        return await self._tokenize_singleton_prompt_async(prompt, params)
504
505
506

    async def tokenize_prompts_async(
        self,
507
        prompts: Sequence[DictPrompt],
508
        params: TokenizeParams,
509
    ) -> list[TokPrompt]:
510
511
512
        return await asyncio.gather(
            *(self.tokenize_prompt_async(prompt, params) for prompt in prompts)
        )
513
514
515
516

    # Step 3: Add extra keys to the prompts
    def _apply_prompt_extras(
        self,
517
        prompts: Sequence[TokPrompt],
518
519
520
521
522
523
        prompt_extras: dict[str, Any] | None,
    ):
        if not prompt_extras:
            return

        for prompt in prompts:
524
            target_prompt = extract_target_prompt(self.model_config, prompt)
525
526
            target_prompt.update(prompt_extras)  # type: ignore[arg-type]

527
528
529
530
    # Step 4: Convert to engine inputs
    def _validate_mm_uuids(
        self,
        mm_data: "MultiModalDataDict",
531
532
        mm_data_items: "MultiModalDataItems",
        mm_uuid_items: "MultiModalUUIDItems",
533
    ) -> None:
534
535
536
        # NOTE: Keys corresponding to `None` in `mm_data` don't appear in
        # `mm_data_items`
        modalities = mm_data.keys() | mm_uuid_items.keys()
537
538

        for modality in modalities:
539
540
            data_items = mm_data_items.get(modality)
            uuid_items = mm_uuid_items.get(modality)
541

542
543
544
545
546
547
            if data_items is None:
                if uuid_items is None:
                    raise ValueError(
                        f"multi_modal_data[{modality!r}] is empty but "
                        f"multi_modal_uuids[{modality!r}] is missing."
                    )
548

549
            elif uuid_items is not None:
550
                if len(data_items) != len(uuid_items):
551
552
553
554
555
556
557
                    raise ValueError(
                        f"If given, multi_modal_uuids[{modality!r}] must have "
                        f"same length as multi_modal_data[{modality!r}], but "
                        f"got {len(uuid_items)} vs {len(data_items)}."
                    )

                for i, item in enumerate(data_items):
558
559
560
561
562
                    if item is None and uuid_items[i] is None:
                        raise ValueError(
                            f"multi_modal_data[{modality!r}][{i}] is empty but "
                            f"multi_modal_uuids[{modality!r}][{i}] is missing."
                        )
563
564
565
566

    def _process_mm_uuids(
        self,
        mm_data: "MultiModalDataDict",
567
568
        mm_data_items: "MultiModalDataItems",
        mm_uuid_items: "MultiModalUUIDItems",
569
570
571
572
573
574
575
576
577
578
579
580
581
582
        mm_req_id: str,
    ):
        model_config = self.model_config

        # NOTE: When users explicitly turn off BOTH prefix caching and input
        # processing caching, no multimodal features or embeddings will be
        # reused across requests, therefore identifying multimodal data items
        # by their content is no longer necessary, and we create uuids with
        # `<mm_req_id>-<modality>-<index>`, overriding even user-provided ones.
        if (
            model_config.multimodal_config
            and model_config.multimodal_config.mm_processor_cache_gb == 0
            and not self.config.cache_config.enable_prefix_caching
        ):
583
            mm_uuid_items = {
584
                modality: [f"{mm_req_id}-{modality}-{i}" for i in range(data_count)]
585
                for modality, data_count in mm_data_items.get_all_counts().items()
586
587
            }

588
        self._validate_mm_uuids(mm_data, mm_data_items, mm_uuid_items)
589

590
        return mm_uuid_items
591
592
593
594
595
596

    # TODO: Remove str and tokenization_kwargs after deprecating InputPreprocessor
    def _process_multimodal(
        self,
        prompt: list[int] | str,
        mm_data: "MultiModalDataDict",
597
        mm_uuids: "MultiModalUUIDDict | None",
598
599
600
        mm_processor_kwargs: Mapping[str, object] | None,
        tokenization_kwargs: dict[str, Any] | None,
    ) -> "MultiModalInputs":
601
        from vllm.multimodal.parse import parse_mm_uuids
602
        from vllm.multimodal.processing import ProcessorInputs as MMProcessorInputs
603

604
        mm_req_id = f"renderer{self.api_process_rank}-mm-{self._mm_req_counter.inc(1)}"
605
606
607

        mm_processor = self.get_mm_processor()

608
609
610
        mm_data_items = mm_processor.info.parse_mm_data(mm_data)
        mm_uuid_items = parse_mm_uuids(mm_uuids)

611
        mm_uuid_items = self._process_mm_uuids(
612
613
            mm_data, mm_data_items, mm_uuid_items, mm_req_id
        )
614

615
616
617
618
619
620
621
622
623
624
625
        mm_processor_inputs = MMProcessorInputs(
            prompt,
            mm_data_items,
            mm_uuid_items,
            hf_processor_mm_kwargs=mm_processor_kwargs or {},
            tokenization_kwargs=tokenization_kwargs or {},
        )
        mm_timing_ctx = self._mm_timing_registry.get(mm_req_id)

        with set_default_torch_num_threads():
            mm_inputs = mm_processor.apply(mm_processor_inputs, mm_timing_ctx)
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723

        self.update_mm_cache_stats()

        return mm_inputs

    def _process_tokens(
        self,
        prompt: TokensPrompt,
    ) -> "TokenInputs | MultiModalInputs":
        prompt_token_ids = prompt["prompt_token_ids"]

        inputs: TokenInputs | MultiModalInputs
        if multi_modal_data := prompt.get("multi_modal_data"):
            inputs = self._process_multimodal(
                prompt_token_ids,
                multi_modal_data,
                mm_processor_kwargs=prompt.get("mm_processor_kwargs"),
                tokenization_kwargs=None,  # Tokenization already done in Step 2
                mm_uuids=prompt.get("multi_modal_uuids"),
            )
        else:
            inputs = token_inputs(prompt_token_ids)

        if prompt_text := prompt.get("prompt"):
            inputs["prompt"] = prompt_text
        if cache_salt := prompt.get("cache_salt"):
            inputs["cache_salt"] = cache_salt

        return inputs

    def _process_embeds(
        self,
        prompt: EmbedsPrompt,
    ) -> EmbedsInputs:
        if not self.model_config.enable_prompt_embeds:
            raise ValueError(
                "You must set `--enable-prompt-embeds` to input `prompt_embeds`."
            )

        prompt_embeds = prompt["prompt_embeds"]

        # prompt_embeds must be (seq_len, hidden_size), but if the user
        # passes in a batch of size 1, i.e. (1, seq_len, hidden_size),
        # we can unambiguously process the intent by squeezing the batch
        # dimension.
        if prompt_embeds.ndim == 3:
            prompt_embeds = prompt_embeds.squeeze(dim=0)

        if prompt_embeds.ndim != 2:
            raise ValueError("prompt_embeds must be of shape (seq_len, hidden_size).")

        # Tensors must be on CPU for serialization between processes
        # in the MsgpackEncoder. Casting to CPU here ensures that there is no
        # hidden device transfer in the critical path of generation.
        prompt_embeds = prompt_embeds.cpu()

        return embeds_inputs(
            prompt_embeds=prompt_embeds,
            cache_salt=prompt.get("cache_salt"),
        )

    def _process_singleton(
        self,
        prompt: SingletonTokPrompt,
    ) -> SingletonInputs:
        if "prompt_embeds" in prompt:
            return self._process_embeds(prompt)  # type: ignore[arg-type]

        return self._process_tokens(prompt)  # type: ignore[arg-type]

    def _process_enc_dec(
        self,
        prompt: EncoderDecoderTokPrompt,
    ) -> EncoderDecoderInputs:
        enc_prompt = prompt["encoder_prompt"]
        dec_prompt = prompt["decoder_prompt"]

        return build_enc_dec_inputs(
            encoder_inputs=self._process_singleton(enc_prompt),
            decoder_inputs=(
                None if dec_prompt is None else self._process_singleton(dec_prompt)
            ),
            decoder_start_token_id=self.get_dec_start_token_id(),
        )

    def process_for_engine(
        self, prompt: TokPrompt, arrival_time: float
    ) -> ProcessorInputs:
        engine_prompt: ProcessorInputs
        if "encoder_prompt" in prompt:
            engine_prompt = self._process_enc_dec(prompt)  # type: ignore[arg-type]
        else:
            engine_prompt = self._process_singleton(prompt)

        engine_prompt["arrival_time"] = arrival_time

        return engine_prompt

724
725
726
727
    # Top-level methods
    def render_cmpl(
        self,
        prompts: Sequence[DictPrompt | bytes],
728
        tok_params: TokenizeParams | None = None,
729
730
731
        *,
        prompt_extras: dict[str, Any] | None = None,
    ):
732
733
        arrival_time = time.time()

734
735
        if tok_params is None:
            tok_params = self.default_cmpl_tok_params
736

737
        dict_prompts = self.render_prompts(prompts)
738
739
740
741
        tok_prompts = self.tokenize_prompts(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

742
        return [self.process_for_engine(prompt, arrival_time) for prompt in tok_prompts]
743
744
745
746

    async def render_cmpl_async(
        self,
        prompts: Sequence[DictPrompt | bytes],
747
        tok_params: TokenizeParams | None = None,
748
749
750
        *,
        prompt_extras: dict[str, Any] | None = None,
    ):
751
752
        arrival_time = time.time()

753
754
        if tok_params is None:
            tok_params = self.default_cmpl_tok_params
755

756
        dict_prompts = await self.render_prompts_async(prompts)
757
758
759
760
        tok_prompts = await self.tokenize_prompts_async(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

761
        return [self.process_for_engine(prompt, arrival_time) for prompt in tok_prompts]
762
763
764
765
766

    def render_chat(
        self,
        conversations: Sequence[list["ChatCompletionMessageParam"]],
        chat_params: ChatParams,
767
        tok_params: TokenizeParams | None = None,
768
769
770
        *,
        prompt_extras: dict[str, Any] | None = None,
    ):
771
772
        arrival_time = time.time()

773
774
775
        if tok_params is None:
            tok_params = self.default_chat_tok_params

776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
        rendered = [
            self.render_messages(conversation, chat_params)
            for conversation in conversations
        ]

        out_conversations = list[list["ConversationMessage"]]()
        dict_prompts = list[DictPrompt]()
        for conv, prompt in rendered:
            out_conversations.append(conv)
            dict_prompts.append(prompt)

        tok_prompts = self.tokenize_prompts(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

791
792
793
794
795
        eng_prompts = [
            self.process_for_engine(prompt, arrival_time) for prompt in tok_prompts
        ]

        return out_conversations, eng_prompts
796
797
798
799
800

    async def render_chat_async(
        self,
        conversations: Sequence[list["ChatCompletionMessageParam"]],
        chat_params: ChatParams,
801
        tok_params: TokenizeParams | None = None,
802
803
804
        *,
        prompt_extras: dict[str, Any] | None = None,
    ):
805
806
        arrival_time = time.time()

807
808
809
        if tok_params is None:
            tok_params = self.default_chat_tok_params

810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
        rendered = [
            self.render_messages_async(conversation, chat_params)
            for conversation in conversations
        ]

        out_conversations = list[list["ConversationMessage"]]()
        dict_prompts = list[DictPrompt]()
        for conv, prompt in await asyncio.gather(*rendered):
            out_conversations.append(conv)
            dict_prompts.append(prompt)

        tok_prompts = await self.tokenize_prompts_async(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

825
826
827
828
829
        eng_prompts = [
            self.process_for_engine(prompt, arrival_time) for prompt in tok_prompts
        ]

        return out_conversations, eng_prompts